update config

config.json

@@ -42,6 +42,7 @@
         "image_size": 224,
         "patch_size": 32,
         "width": 768,
+        "head_width": 64,
         "layers": 12,
         "mlp_ratio": 4.0,
         "ls_init_value": null,

configuration_japanese_clip.py

@@ -10,7 +10,7 @@ class JapaneseCLIPVisionConfig(PretrainedConfig):
         patch_size: int, 
         width: int,
         layers: int,
-        heads: int,
+        head_width: int,
         mlp_ratio: float,
         ls_init_value: float = None,
         attentional_pool: bool = False,
@@ -30,7 +30,8 @@ class JapaneseCLIPVisionConfig(PretrainedConfig):
         self.patch_size = patch_size
         self.width = width
         self.layers = layers
-        self.heads = heads
+        self.head_width = head_width
+        self.heads = width // head_width
         self.mlp_ratio = mlp_ratio
         self.ls_init_value = ls_init_value
         self.attentional_pool = attentional_pool